-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
276 lines (229 loc) · 11.1 KB
/
train.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
import torch
import torch.nn as nn
import time
import argparse
import torchvision.utils as vutils
from model.TLCKDNet import TLCKDNet
from datasets.train_dataset import train_dataloader, val_dataloader
import torch.backends.cudnn as cudnn
from torch.utils.data import DataLoader
import os
from tensorboardX import SummaryWriter
import pytorch_ssim
import pytorch_iou
from utils import S_measure
import logger
from fvcore.nn import FlopCountAnalysis
import torch.nn.functional as F
import numpy as np
import random
# --- Parse hyper-parameters train --- #
parser = argparse.ArgumentParser(description='ORSI-SOD TRAIN')
parser.add_argument('-learning_rate', help='Set the learning rate', default=1e-4, type=float)
parser.add_argument('-min_learning_rate', help='Set minimum learning rate', default=1e-6, type=float)
parser.add_argument('-train_batch_size', help='Set the training batch size', default=12, type=int)
parser.add_argument('-train_epoch', help='Set the training epoch', default=300, type=int)
parser.add_argument('--train_dataset', type=str, default='ORSSD')
parser.add_argument('--data_dir', type=str, default='./data')
parser.add_argument('--checkpoint_dir', type=str, default='./train_results')
parser.add_argument('--log_dir', type=str, default='./results/Log/TLCDNet_ORSSD/')
# --- Parse hyper-parameters val --- #
parser.add_argument('--val_dataset', type=str, default='ORSSD')
parser.add_argument('--predict_result', type=str, default='./output_result/picture/')
parser.add_argument('-val_batch_size', help='Set the testing batch size', default=1, type=int)
parser.add_argument('--imagenet_model', default='./ImgNet', type=str, help='load trained model or not')
parser.add_argument('--snapshot', type=str, default='none', help='pretrained model')
args = parser.parse_args()
args.best_mae = 1
best_epoch = 0
args.best_Sm = 0
# --- set random seeds --- #
def setup_seed(seed):
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
random.seed(seed)
setup_seed(2023)
# --- train --- #
learning_rate = args.learning_rate
min_lr = args.min_learning_rate
train_batch_size = args.train_batch_size
train_epoch= args.train_epoch
train_dataset=os.path.join(args.data_dir, args.train_dataset, 'train')
# --- test --- #
val_dataset = os.path.join(args.data_dir, args.val_dataset, 'val')
predict_result= args.predict_result
val_batch_size=args.val_batch_size
# --- output picture and check point --- #
if not os.path.exists(args.checkpoint_dir):
os.makedirs(args.checkpoint_dir)
if not os.path.exists(args.log_dir):
os.makedirs(args.log_dir)
output_dir=os.path.join(args.checkpoint_dir,'output_result')
logger = logger.create_logger(output_dir=args.log_dir, name="ORSI-SOD")
# --- Gpu device --- #
device_ids = [Id for Id in range(torch.cuda.device_count())]
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# --- cudnn setting --- #
torch.backends.cudnn.enabled = True
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = True
# --- Define the network --- #
ORSISOD_Net = TLCKDNet(args)
i = torch.randn(1,3,256,256)
flops = FlopCountAnalysis(ORSISOD_Net, i)
logger.info(f"MyNetwork parameters: {sum(param.numel() for param in ORSISOD_Net.parameters()) / 1024 ** 2:.2f}M")
logger.info(f"MyNetwork flops: {flops.total() / 1e9:.2f}G")
# --- Build optimizer --- #
warmup_epochs = 3
optimizer = torch.optim.AdamW(ORSISOD_Net.parameters(), lr=learning_rate, weight_decay=0.01)
#scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=[100,250,400], gamma=0.5)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=train_epoch, eta_min=min_lr)
# --- Load training data --- #
dataset = train_dataloader(train_dataset)
train_loader = DataLoader(dataset=dataset, batch_size=train_batch_size, shuffle=True, num_workers=8)
# --- Load val data --- #
val_dataset = val_dataloader(val_dataset)
val_loader = DataLoader(dataset=val_dataset, batch_size=val_batch_size, shuffle=False, num_workers=1)
# --- Multi-GPU --- #
ORSISOD_Net = ORSISOD_Net.to(device)
ORSISOD_Net = torch.nn.DataParallel(ORSISOD_Net, device_ids=device_ids)
# DNet = DNet.to(device)
# DNet= torch.nn.DataParallel(DNet, device_ids=device_ids)
writer = SummaryWriter(os.path.join(args.checkpoint_dir, 'tensorboard'))
# i = torch.randn(1,3,256,256).to(device)
# writer.add_graph(ORSISOD_Net, (i))
# --- Loss function --- #
bce_loss = nn.BCELoss(reduction='mean')
ssim_loss = pytorch_ssim.SSIM(window_size=11,size_average=True)
iou_loss = pytorch_iou.IOU(size_average=True)
def bce_ssim_loss(pred, target):
bce_out = bce_loss(pred,target)
ssim_out = 1 - ssim_loss(pred,target)
# iou_out = iou_loss(pred,target)
loss = bce_out + ssim_out
return loss
# def muti_bce_loss_fusion(d0, d_m, d1, labels_v, edge):
# loss0 = bce_ssim_loss(d0, labels_v)
# loss_m = bce_ssim_loss(d_m, labels_v)
# loss_edge = bce_ssim_loss(d1, edge)
# loss = loss0 + loss_m + 0.3 * loss_edge
#
# return loss, loss0, loss_m, loss_edge
def show_featuremap(model, image,iteration):
# 定义网格
img_grid = vutils.make_grid(image, normalize=True, scale_each=True, nrow=4)
# 绘制原始图像
writer.add_image('raw img', img_grid, global_step=iteration) # j 表示feature map数
model.eval()
for name, layer in model._modules.items():
#print(name, layer)
if not ('ex' in name):
return
image = layer(image)
if 'ex' in name:
x1 = image[0].transpose(0, 1) # C,B, H, W ---> B,C, H, W
x2 = image[1].transpose(0, 1)
x3 = image[2].transpose(0, 1)
x4 = image[3].transpose(0, 1)
img_grid = vutils.make_grid(x1, normalize=True, scale_each=True, nrow=16) # normalize进行归一化处理
img_grid2 = vutils.make_grid(x2, normalize=True, scale_each=True, nrow=8) # normalize进行归一化处理
img_grid3 = vutils.make_grid(x3, normalize=True, scale_each=True, nrow=8) # normalize进行归一化处理
img_grid4 = vutils.make_grid(x4, normalize=True, scale_each=True, nrow=8) # normalize进行归一化处理
writer.add_image('ex1_feature_maps', img_grid, global_step=iteration)
writer.add_image('ex2_feature_maps', img_grid2, global_step=iteration)
writer.add_image('ex3_feature_maps', img_grid3, global_step=iteration)
writer.add_image('ex4_feature_maps', img_grid4, global_step=iteration)
# --- Load the network weight --- #
try:
ORSISOD_Net.load_state_dict(torch.load(os.path.join(args.teacher_model , 'epoch100000.pkl')))
logger.info('--- weight loaded ---')
except:
logger.info('--- no weight loaded ---')
# --- Start training --- #
logger.info('--------------------------------')
logger.info('======== Start training ========')
logger.info('--------------------------------')
# --- Free up graphics memory --- #
torch.cuda.empty_cache()
iteration = 0
for epoch in range(train_epoch):
start_time = time.time()
ORSISOD_Net.train()
logger.info(f"==> Epoch[{epoch}/{train_epoch}]")
for batch_idx, (image, binary, edge) in enumerate(train_loader):
# print(batch_idx)
iteration +=1
image = image.to(device)
binary = binary.to(device)
edge = edge.to(device)
#show_featuremap(MyEnsembleNet, hazy, iteration)
output = ORSISOD_Net(image)
optimizer.zero_grad()
#total_loss, loss0, loss_m, loss_edge = muti_bce_loss_fusion(output, m_output, edge_out, clean, edge)
total_loss =bce_ssim_loss(output, binary)
total_loss.backward()
nn.utils.clip_grad_norm_(ORSISOD_Net.parameters(), max_norm=0.5)
optimizer.step()
writer.add_scalars('training', {'training total loss': total_loss.item()
}, iteration)
scheduler.step()
if epoch % 5 == 0:
logger.info(f"===================> Validation on epoch: {str(epoch)} <===================")
with torch.no_grad():
psnr_list = []
edge_mae_list = []
recon_psnr_list = []
Sm_list = []
ORSISOD_Net.eval()
for batch_idx, (image, binary, edge) in enumerate(val_loader):
image = image.to(device)
binary = binary.to(device)
edge = edge.to(device)
frame_out = ORSISOD_Net(image)
frame_out = F.interpolate(frame_out, size=binary.shape[2:], mode='bilinear', align_corners=True)
# if not os.path.exists(output_dir):
# os.makedirs(output_dir)
# imwrite(frame_out, output_dir +'/' +str(batch_idx) + '.png', range=(0, 1))
# mae = torch.abs(frame_out[0]-clean).mean()
# edge_mae = torch.abs(frame_out[2]-edge).mean()
# Sm = S_measure(frame_out[0], clean).cuda()
# psnr_list.append(mae)
# edge_mae_list.append(edge_mae)
# Sm_list.append(Sm)
mae = torch.abs(frame_out - binary).mean()
#edge_mae = torch.abs(frame_out- edge).mean()
Sm = S_measure(frame_out, binary).cuda()
psnr_list.append(mae)
#edge_mae_list.append(edge_mae)
Sm_list.append(Sm)
avr_mae = sum(psnr_list) / len(psnr_list)
#avr_edge_mae = sum(edge_mae_list) / len(edge_mae_list)
avr_Sm = sum(Sm_list) / len(Sm_list)
# frame_debug = torch.cat((frame_out[0],clean), dim =0)
# edge_img = torch.cat((frame_out[2], edge), dim=0)
# m_s_map = torch.cat((frame_out[1], frame_out[0]), dim=0)
frame_debug = torch.cat((frame_out, binary), dim=0)
#edge_img = torch.cat((frame_out[2], edge), dim=0)
#m_s_map = torch.cat((frame_out[1], frame_out[0]), dim=0)
writer.add_images('salient_image_and_label', frame_debug, epoch)
# writer.add_images('edge_image_and_label', edge_img, epoch)
# writer.add_images('moderate_salient_map_and_pred_map', m_s_map, epoch)
writer.add_scalars('salient_mae_testing', {'test salient_mae':avr_mae.item(),
}, epoch)
# writer.add_scalars('edge_testing', {'test edge_mae': avr_edge_mae.item(),
# }, epoch)
writer.add_scalars('salient_Sm_testing', {'test salient_Sm': avr_Sm.item(),
}, epoch)
#torch.save(MyEnsembleNet.state_dict(), os.path.join(args.model_save_dir,'epoch'+ str(epoch) + '.pkl'))
if args.best_Sm < avr_Sm:
args.best_Sm = avr_Sm
best_epoch = epoch
torch.save(ORSISOD_Net.state_dict(), os.path.join(args.checkpoint_dir,'ORSSD_epoch'+ str(epoch) + '.pkl'))
logger.info(f"epoch: {epoch}\t"
f"salient_mae: {avr_mae.item():.6f}\t"
f"salient_Sm: {avr_Sm.item():.6f}\t"
f"best_epoch: {best_epoch}\t"
f"best_Sm: {args.best_Sm.item():.6f}\t"
f"lr: {optimizer.param_groups[0]['lr']}")
writer.close()